WO2020153047A1 - 電圧制御装置 - Google Patents
電圧制御装置 Download PDFInfo
- Publication number
- WO2020153047A1 WO2020153047A1 PCT/JP2019/049265 JP2019049265W WO2020153047A1 WO 2020153047 A1 WO2020153047 A1 WO 2020153047A1 JP 2019049265 W JP2019049265 W JP 2019049265W WO 2020153047 A1 WO2020153047 A1 WO 2020153047A1
- Authority
- WO
- WIPO (PCT)
- Prior art keywords
- voltage
- neural network
- determination unit
- control device
- inference result
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Ceased
Links
Images
Classifications
-
- G—PHYSICS
- G06—COMPUTING OR CALCULATING; COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F1/00—Details not covered by groups G06F3/00 - G06F13/00 and G06F21/00
- G06F1/26—Power supply means, e.g. regulation thereof
- G06F1/32—Means for saving power
- G06F1/3203—Power management, i.e. event-based initiation of a power-saving mode
- G06F1/3234—Power saving characterised by the action undertaken
- G06F1/3296—Power saving characterised by the action undertaken by lowering the supply or operating voltage
-
- G—PHYSICS
- G06—COMPUTING OR CALCULATING; COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/04—Architecture, e.g. interconnection topology
- G06N3/045—Combinations of networks
-
- G—PHYSICS
- G06—COMPUTING OR CALCULATING; COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/04—Architecture, e.g. interconnection topology
- G06N3/0464—Convolutional networks [CNN, ConvNet]
-
- G—PHYSICS
- G06—COMPUTING OR CALCULATING; COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/06—Physical realisation, i.e. hardware implementation of neural networks, neurons or parts of neurons
- G06N3/063—Physical realisation, i.e. hardware implementation of neural networks, neurons or parts of neurons using electronic means
- G06N3/065—Analogue means
-
- G—PHYSICS
- G06—COMPUTING OR CALCULATING; COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/08—Learning methods
- G06N3/09—Supervised learning
-
- Y—GENERAL TAGGING OF NEW TECHNOLOGICAL DEVELOPMENTS; GENERAL TAGGING OF CROSS-SECTIONAL TECHNOLOGIES SPANNING OVER SEVERAL SECTIONS OF THE IPC; TECHNICAL SUBJECTS COVERED BY FORMER USPC CROSS-REFERENCE ART COLLECTIONS [XRACs] AND DIGESTS
- Y02—TECHNOLOGIES OR APPLICATIONS FOR MITIGATION OR ADAPTATION AGAINST CLIMATE CHANGE
- Y02D—CLIMATE CHANGE MITIGATION TECHNOLOGIES IN INFORMATION AND COMMUNICATION TECHNOLOGIES [ICT], I.E. INFORMATION AND COMMUNICATION TECHNOLOGIES AIMING AT THE REDUCTION OF THEIR OWN ENERGY USE
- Y02D10/00—Energy efficient computing, e.g. low power processors, power management or thermal management
Definitions
- the present technology relates to a voltage control device.
- DNN deep neural networks
- SRAM Static Random Access Memory
- Non-Patent Document 1 the learning neural network learns in an environment in which the voltage supplied to the SRAM of the learning neural network is low and an error in the SRAM is likely to occur.
- this supply voltage becomes lower than a predetermined value, the error rate (Error Rate) of the learning neural network increases. Therefore, the author of this non-patent document 1 sets a threshold value for this error rate and sets an allowable error rate and the lowest supply voltage (limit operating voltage) at which this error rate is reached. Then, while the normal voltage was 0.8 V, the limit operating voltage was 0.5 V or less. The above is described in this nonpatent literature 1.
- Non-Patent Document 1 a learning neural network and an inference neural network are used.
- the coefficient data obtained by the learning by the learning neural network is stored in the SRAM of the inference neural network, and the inference neural network makes an inference. Therefore, it is desirable that the limit operating voltage obtained during learning can be applied during inference.
- the PVT (process, voltage, and temperature) condition during learning and the PVT condition during inference may differ. Therefore, there is a problem that the limit operating voltage obtained during learning cannot be applied to the inference neural network as it is.
- the main purpose of this technology is to reduce power consumption by providing a voltage control device that automatically sets the limit operating voltage.
- the present technology includes a first neural network, a second neural network, an inference result determination unit, and a voltage determination unit, and the first neural network has a function of inferring based on learned information.
- the second neural network has a function of inferring based on unlearned information
- the inference result determining unit includes the correct answer value data held in the inference result determining unit and the first 1 has a function of comparing the inference result data of the neural network to obtain the determination result data
- the voltage determining unit matches the correct answer value data with the inference result data based on the determination result data.
- a voltage signal lower than the voltage supplied to the first neural network and the second neural network is output, and if the correct answer value data and the inference result data do not match, the first neural network and the A voltage controller having a function of outputting the voltage signal higher than the voltage supplied to the second neural network is provided.
- the first neural network may be a duplicate of some or all of the components of the second neural network.
- the first neural network may be formed by a part of the components of the second neural network.
- the first neural network and the second neural network may include at least a storage unit and a plurality of processing elements.
- the voltage determination unit includes at least a voltage difference determination unit and a voltage difference signal conversion unit, and the voltage difference determination unit is supplied to the first neural network based on the determination result data.
- the voltage difference signal converter converts the voltage difference signal, It may have a function of outputting a voltage signal.
- the voltage determination unit based on the determination result data and the voltage, the correct value data and the inference result data match, and the voltage supplied to the first neural network is a predetermined value. If it is higher, a voltage signal lower than the voltage supplied to the first neural network and the second neural network may be output.
- the voltage control device further includes a voltage supply unit, and the voltage supply unit applies the voltage to the first neural network and the second neural network based on the voltage signal output by the voltage determination unit. It may have a function of supplying.
- the voltage supply unit may include at least a DA converter, and the DA converter may have a function of converting the voltage signal that is a digital signal into the voltage that is an analog signal and outputting the voltage.
- the voltage supplied to the first neural network is referred to as a first supply voltage and the voltage supplied to the second neural network is referred to as a second supply voltage. It may have a function of supplying the high second supply voltage to the second neural network.
- the error occurrence rate of the first neural network is within the allowable range, and the lowest voltage supplied to the first neural network is called the limit operating voltage.
- the voltage control device 100 includes a first neural network 110, a second neural network 120, an inference result determination unit 130, and a voltage determination unit 140.
- the first neural network 110 and the second neural network 120 have a function of inferring what the input data is. For example, when image data of a cat is input, the first neural network 110 and the second neural network 120 infer whether or not this image is a cat image.
- the input data may be, for example, voices, symbols, numbers, characters, signals, etc. in addition to images.
- the first neural network 110 has a function of inferring based on already learned information.
- the second neural network 120 has a function of inferring based on unlearned information.
- the first neural network 110 is a replica of some or all of the components of the second neural network 120. As a result, the first neural network 110 and the second neural network 120 can perform the same operation.
- the inference result determination unit 130 has a function of determining whether the inference result of the first neural network 110 is correct or incorrect and obtaining determination result data. More specifically, the inference result determination unit 130 compares the correct answer value data held by the inference result determination unit 130 with the inference result data of the first neural network 110, and outputs the determination result data. .. For example, when the correct answer value data is a cat and the inference result data is also a cat, the determination result data has a correct answer value.
- the voltage determination unit 140 has a function of determining the voltage supplied to the first neural network 110 and the second neural network 120 based on the determination result data output by the inference result determination unit 130. More specifically, when the correct answer value data and the inference result data match, the voltage determination unit 140 outputs a voltage signal lower than the voltage supplied to the first neural network 110 and the second neural network 120. It has a function. On the other hand, when the correct answer value data and the inference result data do not match, the voltage determination unit 140 has a function of outputting a voltage signal higher than the voltage supplied to the first neural network 110 and the second neural network 120. There is.
- the first neural network 110 may be configured with a part of the components of the second neural network 120. That is, the first neural network 110 may operate as the first neural network 110 at some times and as a part of the second neural network 120 at some times. For example, when inferring based on already learned information, the first neural network 110 operates as the first neural network 110, and when inferring based on unlearned information, the first neural network 110: It may operate as part of the second neural network 120.
- the voltage determination unit 140 In order to reduce the power consumption of the first neural network 110 and the second neural network 120, it is effective to lower the voltage supplied to the first neural network 110 and the second neural network 120. Therefore, when the inference result data and the correct answer value data match, the voltage determination unit 140 outputs a voltage signal lower than the voltage supplied to the first neural network 110 and the second neural network 120.
- the first neural network 110 has a function of inferring based on the learned information. Therefore, it is usual that the inference result data of the first neural network 110 and the correct value data match.
- the case where the inference result data and the correct value data do not match is when the voltage supplied to the first neural network 110 becomes lower than the limit operating voltage V L.
- the voltage determination unit 140 outputs a voltage signal higher than the voltage supplied to the first neural network 110 and the second neural network 120.
- the first neural network 110 and the second neural network 120 operate at a voltage near the limit operating voltage V L. Therefore, the power consumption of the first neural network 110 and the second neural network 120 can be reduced.
- the voltage control device 100 can automatically set the limit operating voltage.
- FIG. 3 shows an embodiment in which the voltage control device 100 according to the present technology is embodied.
- the voltage control device 100 according to the present technology includes a first neural network 110, a second neural network 120, an inference result determination unit 130, and a voltage determination unit 140.
- the first neural network 110 can include at least a storage unit 111 and a processing element group 112.
- the processing element group 112 is composed of a plurality of processing elements.
- the storage unit 111 for example, a semiconductor storage element such as SRAM can be used.
- the second neural network 120 can include at least a storage unit 121 and a processing element group 122.
- Inference result judging unit 130 has the inference result determining section and correct value data A C held in the 130, the function to obtain a determination result data by comparing the inference result data of the first neural network 110.
- the inference result determining unit 130 compares the inference and result judging unit 130 correct value data A C held in, the inference result data of the neural network 110 (S1).
- the inference result determination unit 130 sets the value "0" to the determination result flag F which is the determination result data (S2).
- the inference result judging unit 130, the determination result flag F sets the value "1" (S3).
- the voltage determination unit 140 operates based on the value of the determination result flag F.
- FIG. 5 shows a flowchart of the voltage determination unit 140.
- the voltage determination unit 140 outputs a voltage signal lower than the voltage V1 supplied to the neural network 110. Output (S5).
- the voltage determination unit 140 outputs a voltage signal higher than the voltage V1 supplied to the neural network 110 (S6).
- FIGS. 6 and 7 show an embodiment of the first neural network 110 and the inference result determination unit 130.
- FIG. 6 shows a process when the first neural network 110 can correctly infer.
- FIG. 7 shows the processing when the first neural network 110 cannot correctly infer.
- the image data G1 is input to the first neural network 110.
- the image data G1 is a cat image.
- the inference result data by the first neural network 110 is that the image data G1 has a probability of being a cat of 90%, a probability of being a dog of 5%, a probability of being a horse, 2%, a probability of being a bird, and a sheep.
- the probability is 1%.
- the voltage V1 supplied to the first neural network 110 is higher than the limit operating voltage V L. Therefore, the first neural network 110 can make inference correctly.
- the binarization unit 131 included in the inference result determination unit 130 binarizes the inference result data by the first neural network 110 into “0” or “1” using a predetermined value, for example, 50% as a threshold. When the probability in the inference result data is higher than 50%, the binarization unit 131 outputs the value “1”. If the probability in the inference result data is 50% or less, the binarization unit 131 outputs the value “0”. The threshold does not have to be 50%.
- the probability that the image data is a cat is 90%, which is higher than the threshold value of 50%, so the binarization unit 131 sets the value of the cat item to “1”. For other items (dogs, horses, birds, and sheep), the probability is 50% or less, so the binarization unit 131 sets the values of these items to “0”.
- the correct answer is a cat because the image data is the image data of a cat. Therefore, the correct value data A C, the value of the cat of the item has become a "1", other items (dogs, horses, birds, and sheep) the value is and has a "0".
- the correct answer comparing unit 132 included in the inference result determining unit 130 compares the correct answer value data A C and reasoning result data.
- the cat item, the value of the inference result data is "1", the value of the correct value data A C is “1”, "1” is set to the determination value.
- the value of the inference result data is "0”, the value of the correct value data A C is "0”, "1” is set to the determination value ..
- the inference result judgment unit 130 sets the judgment result flag F to the value "0".
- the image data G1 is a cat image.
- the inference result data obtained by the first neural network 110 has a probability that the image data is a cat is 45%, a dog is 51%, a horse is 1%, a bird is 2%, and a sheep is a probability. Is 1%.
- the voltage V1 supplied to the first neural network 110 is lower than the limit operating voltage V L. Therefore, the first neural network 110 cannot correctly make inference.
- the probability that the image data is a dog is 51%, which is higher than the threshold value 50%, so the binarization unit 131 sets the value of the dog item to “1”. For other items (cat, horse, bird, and sheep), the probability is 50% or less, so the binarization unit 131 sets the values of these items to “0”.
- the correct answer comparing unit 132 compares the correct answer value data A C and reasoning result data.
- the cat item the value of the inference result data is "0", the value of the correct value data A C is “1”, "0” is set to the determination value.
- the value of the inference result data is "1”, the value of the correct value data A C is "0", "0” is set to the determination value.
- the value of the inference result data is "0"
- the value of the correct value data A C is "0"
- "1" is set to the determination value.
- the inference result judgment unit 130 sets the judgment result flag F to the value “1”. To set.
- the subsequent voltage determination unit 140 operates based on the determination result flag F.
- the voltage determination unit 140 includes at least a voltage difference determination unit 141 and a voltage difference signal conversion unit 142.
- the voltage difference determination unit 141 determines the voltage difference between the voltage supplied to the first neural network 110 and the changed voltage based on the determination result data, and outputs the voltage difference signal to the voltage difference signal conversion unit 142. It has a function to output.
- the voltage difference determination unit 141 holds the ratio of the voltage difference when lowering the supply voltage V1 to the first neural network and the voltage difference when increasing the supply voltage V1.
- the ratio of the voltage difference when lowering the supply voltage V1 and the voltage difference when increasing the supply voltage V1 can be ⁇ :(1 ⁇ ).
- ⁇ can be a value greater than 0 and less than 0.5.
- the voltage difference signal conversion unit 142 has a function of converting the voltage difference signal and outputting the voltage signal.
- the voltage difference signal conversion unit 142 can include, for example, an integrator.
- FIG. 8 shows the supply voltage V1 to the first neural network, the limit operating voltage V L , the value of the determination result flag F, and the time T.
- the value of the judgment result flag F is indicated by 0 and 1.
- a flat low value is 0 and a high flat value is 1.
- the supply voltage V1 When the value of the determination result flag F is 0, the supply voltage V1 is low. When the supply voltage V1 gradually decreases and the supply voltage V1 reaches the limit operating voltage VL at time t1, the value of the determination result flag F becomes 1 and the supply voltage V1 is high.
- the ratio between the voltage difference when the supply voltage V1 is lowered and the voltage difference when the supply voltage is raised is not limited to ⁇ :1- ⁇ , and may be, for example, ⁇ :1.
- FIG. 9 shows the supply voltage V1 to the first neural network 110, the limit operating voltage VL , the value of the determination result flag F, and the time T.
- the supply voltage V1 gradually decreases, but when the supply voltage V1 becomes close to the limit operating voltage VL at time t3, the supply voltage V1 becomes lower than that. Not not. As a result, at a time after the time t3, the first neural network 110 does not make a false inference although the supply voltage is low. This state may be used when the first neural network 110 is not actively operating and can operate even at a low supply voltage.
- FIG. 10 shows a flowchart of the voltage determination unit 140 at this time.
- the voltage determination unit 140 outputs a voltage signal lower than the voltage supplied to the first neural network and the second neural network (S9). ..
- the voltage determination unit 140 determines that the voltage V1 supplied to the neural network 110 A high voltage signal is output (S10).
- the voltage difference signal conversion unit 142 has a function of converting the voltage difference signal output from the voltage difference determination unit 141 and outputting the voltage signal.
- the voltage difference signal conversion unit 142 multiplies the voltage difference output from the voltage difference determination unit 141 by ⁇ , for example.
- ⁇ can be set to a value larger than 0 and much smaller than 1 for example.
- FIG. 11 shows another embodiment of the voltage control device 100 according to the present technology. Detailed description of the overlapping portions between FIG. 11 and FIG. 3 will be omitted.
- the voltage control device 100 can further include a voltage supply unit 150.
- the voltage supply unit 150 has a function of supplying a voltage to the first neural network 110 and the second neural network 120 based on the voltage signal output by the voltage determination unit 140.
- the voltage supply unit 150 can include at least the DA converter 151.
- the DA converter 151 can convert a digital signal into an analog signal.
- the voltage signal output by the voltage determination unit 140 is a digital signal.
- the supply voltage to the first neural network 110 and the second neural network 120 is an analog signal. Therefore, conversion processing by the DA converter 151 is required.
- the voltage supply unit 150 can further include a buffer 152. If the output voltage of the DA converter 151 is directly supplied to the first neural network 110 or the like, the load of the DA converter 151 or the like becomes high. Therefore, the buffer 152 converts the output voltage of the DA converter 151 and supplies it to the first neural network 110 and the like.
- the buffer 152 for example, a low dropout linear regulator (LDO) or a DC/DC converter can be used.
- FIG. 12 shows another embodiment in which the voltage control device 100 according to the present technology is embodied. Detailed description of the overlapping portions between FIG. 12 and FIG. 11 will be omitted.
- the voltage supplied to the first neural network 110 is the first supply voltage V1.
- the voltage supplied to the second neural network 120 is referred to as a second supply voltage V2.
- the voltage difference signal conversion unit 142 converts the voltage difference signal output by the voltage difference determination unit 141 to generate a voltage signal corresponding to the first supply voltage V1 and a voltage corresponding to the second supply voltage V2 higher than the voltage signal. It has a function of outputting a signal.
- the voltage difference ⁇ V between the first supply voltage V1 and the second supply voltage V2 depends on the fluctuation of noise, but may be, for example, 20 mV or 50 mV.
- the voltage supply unit 150 has a function of supplying the first supply voltage V1 to the first neural network 110 and supplying the second supply voltage V2 higher than the first supply voltage V1 to the second neural network 120.
- FIG. 13 shows the first supply voltage V1, the second supply voltage V2, the limit operating voltage VL , the value of the determination result flag F, and the time T.
- the second supply voltage V2 is slightly higher than the first supply voltage V1. As a result, when the first supply voltage V1 reaches the limit operating voltage V L , the first neural network 110 generates an error, but the second supply voltage V2 does not reach the limit operating voltage V L. The neural network 120 does not generate an error.
- the target for which the voltage control device 100 according to the present technology controls the voltage is not limited to the neural network.
- a first neural network A second neural network, An inference result determination unit, And a voltage determination unit
- the first neural network has a function of inferring based on learned information
- the second neural network has a function of inferring based on unlearned information
- the inference result determination unit has a function of comparing the correct value data held in the inference result determination unit with the inference result data of the first neural network to obtain determination result data, When the correct value data and the inference result data match based on the determination result data, the voltage determination unit outputs a voltage signal lower than the voltage supplied to the first neural network and the second neural network.
- the voltage control device [2] The voltage control device according to [1], wherein the first neural network is a copy of some or all of the components of the second neural network. [3] The voltage control device according to [1] or [2], wherein the first neural network is configured by a part of the constituent elements of the second neural network. [4] The voltage according to any one of [1] to [3], wherein the first neural network and the second neural network include at least a storage unit and a plurality of processing elements. Control device.
- the voltage determining unit includes at least a voltage difference determining unit and a voltage difference signal converting unit, and the voltage difference determining unit is based on the determination result data and the first neural network. Has a function of determining a voltage difference between the voltage being supplied to the voltage and the changed voltage, and outputting a voltage difference signal, wherein the voltage difference signal conversion unit converts the voltage difference signal.
- the voltage control device according to any one of [1] to [4], which has a function of outputting the voltage signal.
- the voltage determining unit determines, based on the determination result data and the voltage, that the correct answer value data and the inference result data match, and that the voltage supplied to the first neural network is The voltage according to any one of [1] to [5], which outputs a voltage signal lower than the voltage supplied to the first neural network and the second neural network when the voltage is higher than a predetermined value.
- Control device [7] A function of further comprising a voltage supply unit, wherein the voltage supply unit supplies the voltage to the first neural network and the second neural network based on the voltage signal output by the voltage determination unit The voltage control device according to any one of [1] to [6].
- the voltage supply unit has at least a DA converter, and the DA converter has a function of converting the voltage signal that is a digital signal into the voltage that is an analog signal and outputting the voltage.
- the voltage control device according to [7].
- the voltage supplied to the first neural network is the first supply voltage
- the voltage supplied to the second neural network is the second supply voltage
- the voltage supply unit is the first
- the voltage control device according to [7] or [8], which has a function of supplying the second supply voltage higher than the supply voltage to the second neural network.
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Physics & Mathematics (AREA)
- General Engineering & Computer Science (AREA)
- General Physics & Mathematics (AREA)
- Health & Medical Sciences (AREA)
- Life Sciences & Earth Sciences (AREA)
- Biomedical Technology (AREA)
- Biophysics (AREA)
- Artificial Intelligence (AREA)
- Computational Linguistics (AREA)
- Data Mining & Analysis (AREA)
- Evolutionary Computation (AREA)
- General Health & Medical Sciences (AREA)
- Molecular Biology (AREA)
- Computing Systems (AREA)
- Mathematical Physics (AREA)
- Software Systems (AREA)
- Neurology (AREA)
- Power Sources (AREA)
- Control Of Voltage And Current In General (AREA)
Abstract
Description
前記第1ニューラルネットワークが、前記第2ニューラルネットワークの一部又は全部の構成要素が複製されたものであってもよい。
前記第1ニューラルネットワークが、前記第2ニューラルネットワークの構成要素の一部で構成されていてもよい。
前記第1ニューラルネットワーク及び前記第2ニューラルネットワークが、少なくとも、記憶部と、複数のプロセッシングエレメントと、を有していてもよい。
前記電圧決定部が、少なくとも、電圧差決定部と、電圧差信号変換部と、を有しており、前記電圧差決定部が、前記判定結果データに基づいて、前記第1ニューラルネットワークに供給されている前記電圧と、変更後の前記電圧との電圧差を決定し、電圧差信号を出力する機能を有しており、前記電圧差信号変換部が、前記電圧差信号を変換して、前記電圧信号を出力する機能を有していてもよい。
前記電圧決定部が、前記判定結果データと前記電圧に基づいて、前記正解値データと前記推論結果データが一致しており、かつ、前記第1ニューラルネットワークに供給されている前記電圧が所定の値より高い場合は、前記第1ニューラルネットワーク及び前記第2ニューラルネットワークに供給されている前記電圧より低い電圧信号を出力してもよい。
前記電圧制御装置が、電圧供給部をさらに備えており、前記電圧供給部が、前記電圧決定部が出力した前記電圧信号に基づいて、前記電圧を前記第1ニューラルネットワーク及び前記第2ニューラルネットワークに供給する機能を有していてもよい。
前記電圧供給部が、少なくともDAコンバーターを有しており、前記DAコンバーターが、デジタル信号である前記電圧信号を、アナログ信号である前記電圧に変換して出力する機能を有していてもよい。
前記第1ニューラルネットワークに供給されている前記電圧を第1供給電圧とし、前記第2ニューラルネットワークに供給されている前記電圧を第2供給電圧とし、前記電圧供給部が、前記第1供給電圧より高い前記第2供給電圧を、前記第2ニューラルネットワークに供給する機能を有していてもよい。
1.本技術に係る第1の実施形態(基本構成)
2.本技術に係る第2の実施形態(電圧供給部)
3.本技術に係る第3の実施形態(異なる供給電圧)
[1]第1ニューラルネットワークと、
第2ニューラルネットワークと、
推論結果判定部と、
電圧決定部と、を備えており、
前記第1ニューラルネットワークが、学習済みの情報に基づいて推論する機能を有しており、
前記第2ニューラルネットワークが、未学習の情報に基づいて推論する機能を有しており、
前記推論結果判定部が、前記推論結果判定部に保持された正解値データと前記第1ニューラルネットワークの推論結果データとを比較して判定結果データを得る機能を有しており、
前記電圧決定部が、前記判定結果データに基づいて、前記正解値データと前記推論結果データが一致した場合は、前記第1ニューラルネットワーク及び前記第2ニューラルネットワークに供給されている電圧より低い電圧信号を出力し、前記正解値データと前記推論結果データが一致しない場合は、前記第1ニューラルネットワーク及び前記第2ニューラルネットワークに供給されている前記電圧より高い前記電圧信号を出力する機能を有している、
電圧制御装置。
[2]前記第1ニューラルネットワークが、前記第2ニューラルネットワークの一部又は全部の構成要素が複製されたものである、[1]に記載の電圧制御装置。
[3]前記第1ニューラルネットワークが、前記第2ニューラルネットワークの構成要素の一部で構成されている、[1]又は[2]に記載の電圧制御装置。
[4]前記第1ニューラルネットワーク及び前記第2ニューラルネットワークが、少なくとも、記憶部と、複数のプロセッシングエレメントと、を有している、[1]~[3]のいずれか一つに記載の電圧制御装置。
[5]前記電圧決定部が、少なくとも、電圧差決定部と、電圧差信号変換部と、を有しており、前記電圧差決定部が、前記判定結果データに基づいて、前記第1ニューラルネットワークに供給されている前記電圧と、変更後の前記電圧との電圧差を決定し、電圧差信号を出力する機能を有しており、前記電圧差信号変換部が、前記電圧差信号を変換して、前記電圧信号を出力する機能を有している、[1]~[4]のいずれか一つに記載の電圧制御装置。
[6]前記電圧決定部が、前記判定結果データと前記電圧に基づいて、前記正解値データと前記推論結果データが一致しており、かつ、前記第1ニューラルネットワークに供給されている前記電圧が所定の値より高い場合は、前記第1ニューラルネットワーク及び前記第2ニューラルネットワークに供給されている前記電圧より低い電圧信号を出力する、[1]~[5]のいずれか一つに記載の電圧制御装置。
[7]電圧供給部をさらに備えており、前記電圧供給部が、前記電圧決定部が出力した前記電圧信号に基づいて、前記電圧を前記第1ニューラルネットワーク及び前記第2ニューラルネットワークに供給する機能を有している、[1]~[6]のいずれか一つに記載の電圧制御装置。
[8]前記電圧供給部が、少なくともDAコンバーターを有しており、前記DAコンバーターが、デジタル信号である前記電圧信号を、アナログ信号である前記電圧に変換して出力する機能を有している、[7]に記載の電圧制御装置。
[9]前記第1ニューラルネットワークに供給されている前記電圧を第1供給電圧とし、前記第2ニューラルネットワークに供給されている前記電圧を第2供給電圧とし、前記電圧供給部が、前記第1供給電圧より高い前記第2供給電圧を、前記第2ニューラルネットワークに供給する機能を有している、[7]又は[8]に記載の電圧制御装置。
110 第1ニューラルネットワーク
120 第2ニューラルネットワーク
130 推論結果判定部
140 電圧決定部
141 電圧差決定部
142 電圧差信号変換部
150 電圧供給部
151 DAコンバーター
V1 第1供給電圧
V2 第2供給電圧
VL 限界動作電圧
F 判定結果フラグ
T 時間
AC 正解値データ
Claims (9)
- 第1ニューラルネットワークと、
第2ニューラルネットワークと、
推論結果判定部と、
電圧決定部と、を備えており、
前記第1ニューラルネットワークが、学習済みの情報に基づいて推論する機能を有しており、
前記第2ニューラルネットワークが、未学習の情報に基づいて推論する機能を有しており、
前記推論結果判定部が、前記推論結果判定部に保持された正解値データと前記第1ニューラルネットワークの推論結果データとを比較して判定結果データを得る機能を有しており、
前記電圧決定部が、前記判定結果データに基づいて、前記正解値データと前記推論結果データが一致した場合は、前記第1ニューラルネットワーク及び前記第2ニューラルネットワークに供給されている電圧より低い電圧信号を出力し、前記正解値データと前記推論結果データが一致しない場合は、前記第1ニューラルネットワーク及び前記第2ニューラルネットワークに供給されている前記電圧より高い前記電圧信号を出力する機能を有している、
電圧制御装置。 - 前記第1ニューラルネットワークが、前記第2ニューラルネットワークの一部又は全部の構成要素が複製されたものである、
請求項1に記載の電圧制御装置。 - 前記第1ニューラルネットワークが、前記第2ニューラルネットワークの構成要素の一部で構成されている、
請求項1に記載の電圧制御装置。 - 前記第1ニューラルネットワーク及び前記第2ニューラルネットワークが、少なくとも、
記憶部と、
複数のプロセッシングエレメントと、を有している、
請求項1に記載の電圧制御装置。 - 前記電圧決定部が、少なくとも、
電圧差決定部と、
電圧差信号変換部と、を有しており、
前記電圧差決定部が、前記判定結果データに基づいて、前記第1ニューラルネットワークに供給されている前記電圧と、変更後の前記電圧との電圧差を決定し、電圧差信号を出力する機能を有しており、
前記電圧差信号変換部が、前記電圧差信号を変換して、前記電圧信号を出力する機能を有している、
請求項1に記載の電圧制御装置。 - 前記電圧決定部が、前記判定結果データと前記電圧に基づいて、前記正解値データと前記推論結果データが一致しており、かつ、前記第1ニューラルネットワークに供給されている前記電圧が所定の値より高い場合は、前記第1ニューラルネットワーク及び前記第2ニューラルネットワークに供給されている前記電圧より低い電圧信号を出力する、
請求項1に記載の電圧制御装置。 - 電圧供給部をさらに備えており、
前記電圧供給部が、前記電圧決定部が出力した前記電圧信号に基づいて、前記電圧を前記第1ニューラルネットワーク及び前記第2ニューラルネットワークに供給する機能を有している、
請求項1に記載の電圧制御装置。 - 前記電圧供給部が、少なくともDAコンバーターを有しており、
前記DAコンバーターが、デジタル信号である前記電圧信号を、アナログ信号である前記電圧に変換して出力する機能を有している、
請求項7に記載の電圧制御装置。 - 前記第1ニューラルネットワークに供給されている前記電圧を第1供給電圧とし、
前記第2ニューラルネットワークに供給されている前記電圧を第2供給電圧とし、
前記電圧供給部が、前記第1供給電圧より高い前記第2供給電圧を、前記第2ニューラルネットワークに供給する機能を有している、
請求項7に記載の電圧制御装置。
Priority Applications (4)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| JP2020567425A JP7404276B2 (ja) | 2019-01-24 | 2019-12-17 | 電圧制御装置 |
| DE112019006731.1T DE112019006731T5 (de) | 2019-01-24 | 2019-12-17 | Spannungssteuerungsvorrichtung |
| US17/422,974 US12298837B2 (en) | 2019-01-24 | 2019-12-17 | Voltage control device |
| CN201980087476.4A CN113316752B (zh) | 2019-01-24 | 2019-12-17 | 电压控制装置 |
Applications Claiming Priority (2)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| JP2019009847 | 2019-01-24 | ||
| JP2019-009847 | 2019-01-24 |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| WO2020153047A1 true WO2020153047A1 (ja) | 2020-07-30 |
Family
ID=71735673
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| PCT/JP2019/049265 Ceased WO2020153047A1 (ja) | 2019-01-24 | 2019-12-17 | 電圧制御装置 |
Country Status (5)
| Country | Link |
|---|---|
| US (1) | US12298837B2 (ja) |
| JP (1) | JP7404276B2 (ja) |
| CN (1) | CN113316752B (ja) |
| DE (1) | DE112019006731T5 (ja) |
| WO (1) | WO2020153047A1 (ja) |
Families Citing this family (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US20230079229A1 (en) * | 2021-09-10 | 2023-03-16 | Maxim Integrated Products, Inc. | Power modulation using dynamic voltage and frequency scaling |
Citations (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| JP2013050841A (ja) * | 2011-08-31 | 2013-03-14 | Nec Corp | メモリ電圧制御装置およびメモリ電圧制御方法 |
| JP2018005297A (ja) * | 2016-06-27 | 2018-01-11 | 富士通株式会社 | ニューラルネットワーク装置及びニューラルネットワーク装置の制御方法 |
Family Cites Families (9)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| JPH0643948A (ja) | 1992-07-27 | 1994-02-18 | Tohoku Electric Power Co Inc | 電圧無効電力制御装置 |
| EP1444600A1 (en) | 2001-11-16 | 2004-08-11 | Yuan Yan Chen | Pausible neural network with supervised and unsupervised cluster analysis |
| CN100570388C (zh) * | 2003-12-18 | 2009-12-16 | 株式会社Lg化学 | 使用神经网络评估电池的充电状态的设备和方法 |
| JP6501146B2 (ja) * | 2014-03-18 | 2019-04-17 | パナソニックIpマネジメント株式会社 | ニューラルネットワーク回路およびその学習方法 |
| CN106875940B (zh) | 2017-03-06 | 2020-08-14 | 吉林省盛创科技有限公司 | 一种基于神经网络的机器自学习构建知识图谱训练方法 |
| CN107239852A (zh) * | 2017-05-05 | 2017-10-10 | 南京邮电大学 | 一种基于深度学习的电量消耗预测方法 |
| CN107562113B (zh) | 2017-08-22 | 2019-04-05 | 电子科技大学 | 一种基于神经网络的低线性调整率基准电路及产生方法 |
| US20190087712A1 (en) * | 2017-09-18 | 2019-03-21 | Qualcomm Incorporated | Neural Network Co-Processing |
| CN108009640B (zh) * | 2017-12-25 | 2020-04-28 | 清华大学 | 基于忆阻器的神经网络的训练装置及其训练方法 |
-
2019
- 2019-12-17 WO PCT/JP2019/049265 patent/WO2020153047A1/ja not_active Ceased
- 2019-12-17 US US17/422,974 patent/US12298837B2/en active Active
- 2019-12-17 CN CN201980087476.4A patent/CN113316752B/zh active Active
- 2019-12-17 DE DE112019006731.1T patent/DE112019006731T5/de active Pending
- 2019-12-17 JP JP2020567425A patent/JP7404276B2/ja active Active
Patent Citations (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| JP2013050841A (ja) * | 2011-08-31 | 2013-03-14 | Nec Corp | メモリ電圧制御装置およびメモリ電圧制御方法 |
| JP2018005297A (ja) * | 2016-06-27 | 2018-01-11 | 富士通株式会社 | ニューラルネットワーク装置及びニューラルネットワーク装置の制御方法 |
Also Published As
| Publication number | Publication date |
|---|---|
| US12298837B2 (en) | 2025-05-13 |
| DE112019006731T5 (de) | 2021-11-18 |
| CN113316752A (zh) | 2021-08-27 |
| CN113316752B (zh) | 2024-09-06 |
| JPWO2020153047A1 (ja) | 2021-12-02 |
| JP7404276B2 (ja) | 2023-12-25 |
| US20220075444A1 (en) | 2022-03-10 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| Goodfellow et al. | Regularization for deep learning | |
| Lanzi et al. | Toward optimal classifier system performance in non-Markov environments | |
| US11262778B2 (en) | Reference voltage generation | |
| US20240211799A1 (en) | System and method for classifying data samples | |
| Sacramento et al. | Energy efficient sparse connectivity from imbalanced synaptic plasticity rules | |
| WO2020153047A1 (ja) | 電圧制御装置 | |
| Zhang et al. | Robust associative learning is sufficient to explain the structural and dynamical properties of local cortical circuits | |
| CN112559738A (zh) | 基于自适应不确定性正则化的情感分类持续学习方法 | |
| Khajeh et al. | Sparse balance: Excitatory-inhibitory networks with small bias currents and broadly distributed synaptic weights | |
| WO2024197399A1 (en) | System and method for evaluating generative artificial intelligence models | |
| Bauso et al. | Mean-field games for marriage | |
| Luo et al. | Robust control and synchronization of 3‐D uncertain fractional‐order chaotic systems with external disturbances via adding one power integrator control | |
| US7561484B2 (en) | Reference-free sampled sensing | |
| CN119621473B (zh) | 一种液冷服务器的温度控制预测方法、装置、设备及介质 | |
| Fan et al. | A hybrid model of extreme learning machine based on bat and cuckoo search algorithm for regression and multiclass classification | |
| Pujara et al. | Using classifier cascades for scalable e-mail classification | |
| Amir et al. | Free trade versus autarky under asymmetric Cournot oligopoly | |
| Guerrero et al. | Comparison of the effect of parameter adaptation in bio-inspired CS algorithm using type-2 fuzzy logic | |
| JP2023137116A (ja) | 情報処理装置、情報処理方法及び情報処理プログラム | |
| Wang et al. | Decision Rules Acquisition for Inconsistent Disjunctive Set‐Valued Ordered Decision Information Systems | |
| Vidal-Saez et al. | Structural determinants of soft memory in recurrent biological networks | |
| CN115269770A (zh) | 基于方面观点词情感对齐的跨领域情感分析方法及装置 | |
| Bach | Are all kernels cursed | |
| US20100179671A1 (en) | System and method for embedding emotion in logic systems | |
| US12353261B2 (en) | Neural-network-based power management for neural network loads |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| 121 | Ep: the epo has been informed by wipo that ep was designated in this application |
Ref document number: 19911965 Country of ref document: EP Kind code of ref document: A1 |
|
| ENP | Entry into the national phase |
Ref document number: 2020567425 Country of ref document: JP Kind code of ref document: A |
|
| 122 | Ep: pct application non-entry in european phase |
Ref document number: 19911965 Country of ref document: EP Kind code of ref document: A1 |
|
| WWG | Wipo information: grant in national office |
Ref document number: 17422974 Country of ref document: US |